adk-architecture

What problem does it solve?

ADK developers need a reliable mental model of how the runtime executes nodes inside workflows, how events drive state reconstruction, and how resumption, routing, and LLM context orchestration work across checkpoints and branches.

Core Features & Use Cases

• Node contract clarity: Explain how BaseNode.run() normalizes yields into Event objects, including routing (ctx.route), outputs (ctx.output), interrupts (long_running_tool_ids), and state deltas (ctx.state).
• Execution flow and responsibilities: Describe how Runner, Workflow, and NodeRunner split responsibilities to avoid deadlocks and to correctly manage concurrency, dynamic scheduling, and persistence/streaming.
• Checkpointing and resume semantics: Provide an actionable guide for HITL lifecycles, run_id reuse, rerun_on_resume behaviors (Workflow vs leaf nodes), and how partial resume_inputs affect control flow.
• LLM context orchestration from events: Cover how the framework constructs an orchestrated LLM view from immutable events, including task delegation translation, branch isolation, and history trimming/compaction to prevent context pollution.

Quick Start

Ask the assistant: "Explain how ADK resumes a workflow after an interrupt, focusing on event flow, context reconstruction, and the difference between ctx.output, ctx.route, and ctx.interrupt_ids."